Izumi.SICK 0.1.6

SICK: Streams of Independent Constant Keys

SICK is an approach to handle JSON-like structures and various libraries implementing it.

SICK allows you to achieve the following:

1. Store JSON-like data in efficient indexed binary form
2. Avoid reading and parsing whole JSON files and access only the data you need just in time
3. Store multiple JSON-like structures in one deduplicating storage
4. Implement perfect streaming parsers for JSON-like data
5. Efficiently stream updates for JSON-like data

The tradeoff for these benefits is somehow more complicated and less efficient encoder.

The problem

JSON has a Type-2 grammar and requires a pushdown automaton to parse it. So, it's not possible to implement efficient streaming parser for JSON. Just imagine a huge hierarchy of nested JSON objects: you won't be able to finish parsing the top-level object until you process the whole file.

JSON is frequently used to store and transfer large amounts of data and these transfers tend to grow over time. Just imagine a typical JSON config file for a large enterprise product.

The non-streaming nature of almost all the JSON parsers requires a lot of work to be done every time you need to deserialize a huge chunk of JSON data: you need to read it from disk, parse it in memory into an AST representation, and, usually, map raw JSON tree to object instances. Even if you use token streams and know the type of your object ahead of time you still have to deal with the Type-2 grammar.

This may be very inefficient and causes unnecessary delays, pauses, CPU activity and memory consumption spikes.

The idea

Let's assume that we have a small JSON:

[
    {"some key": "some value"},
    {"some key": "some value"},
    {"some value": "some key"},
]

Let's build a table for every unique value in our JSON :

We just built a flattened and deduplicated version of our initial JSON structure.

Streaming

Such representation allows us to do many different things, for example we may stream our table:

string:0 = "some key"
string:1 = "some value"

object:0.size = 2
object:0[string:0] = string:1
object:1[string:1] = string:0

array:0.size = 2
array:0[0] = object:0
array:0[1] = object:1

string:2 = "file.json"

root:0=array.0,string:2

This particular encoding is inefficient but it's streamable and, moreover, we can add removal message into it thus supporting arbitrary updates:

array:0[0] = object:1
array:0[1] = remove

There is an interesting observation: when a stream does not contain removal entries it can be safely reordered. Unfortunately, in some usecases the receiver still may need to accumulate the entries in a buffer until it can sort them out.

Binary storage

We may note that the only complex data structures in our "Value" column are lists and (type, index) pairs. Let's call such pairs "references".

A reference can be represented as a pair of integers, so it would have a fixed byte length.

A list of references can be represented as an integer storing list length followed by all the references in their binary form. Let's note that such binary structure is indexed, once we know the index of an element we want to access we can do it immediately.

A list of any fixed-size scalar values can be represented the same way.

A list of variable-size values (e.g. a list of strings) can be represented the following way:

  {strings count}{list of string offsets}{all the strings concatenated}

So, ["a", "bb", "ccc"] would become something like 3 0 2 3 a b bb ccc without spaces.

An important fact is that this encoding is indexed too and it can be reused to store any lists of variable-length data.

Additional capabilities over JSON

SICK encoding follows compositional principles of JSON (a set primitive types plus lists and dictionaries), though it is more powerful: it has "reference" type and allows you to encode custom types.

(1) It's easy to note that our table may store circular references, something JSON can't do natively:

This may be convenient in some complex cases.

(2) Also we may note, that we may happily store multiple json files in one table and have full deduplication over their content. We just need to introduce a separate attribute (is root) storing either nothing or the name of our "root entry" (JSON file).

In real implementation it's more convenient to just create a separate "root" type, the value of a root type should always be a reference to its name and a reference to the actual JSON value we encoded:

(3) We may encode custom scalar data types (e.g. timestamps) natively just by introducing new type tags.

(4) We may even store polymorphic types by introducing new type tags or even new type references.

Implementation

Currently we provide C# and Scala implementations of SICK indexed binary JSON storage. Currently the code in this repository has no streaming capabilities. That may change in the future. It's not a hard problem to add streaming support, your contributions are welcome.

Efficient Binary Indexed Storage

TODO

Supported types

A type marker is represented as a single-byte unsigned integer. The possible values are:

References

TODO

Lists

TODO

Array entries

Array entries are just references.

Object entries

TODO

Object entry skip list and KHash

TODO

Value tables

TODO

Limitations

Current implementation has the following limitations:

1. Maximum object size: 65534 keys
2. The order of object keys is not preserved
3. Maximum amount of array elements: 2^32
4. Maximum amount of unique values of the same type: 2^32

These limitations may be lifted by using more bytes to store offset pointers and counts on binary level. Though it's hard to imagine a real application which would need that.